Maintenance routines on ships today follow either a reactive maintenance (RM) or preventive maintenance (PvM) approach. RM can be regarded as post-failure repair, which might create large costs. PvM uses predetermined maintenance intervals, which often involves unnecessary maintenance. Recently, prognostics and health management (PHM) has emerged as a potential way to develop an ideal maintenance policy. PHM aims to provide optimal maintenance schedule through the use of sensor measurement for fault detection and fault prognostics, among which fault detection is the first and fundamental action. In this paper, a long-short term memory based variational autoencoder (LSTM-VAE) is proposed for fault detection of maritime components onboard. It is a semi-supervised approach that requires only fault-free data for training. Therefore, it is widely applicable in the maritime industry since operational data in normal conditions already exists. Real-world operation data collected from a diesel engine on the research vessel (RV) Gunnerus is used to validate the method. Results show that the LSTM-VAE can detect the fault accurately.

While automatic controllers are frequently used during transit operations and low-speed maneuvering of ships, ship operators typically perform docking maneuvers. This task is more or less challenging depending on factors such as local environment disturbances, number of nearby vessels, and the speed of the ship as it docks. This paper proposes a tool for onboard support that offers position predictions based on an integration of a supervised machine learning (ML) model of the ship into the ship dynamic model. The ML model is applied as a compensator of the unmodelled behaviour or inaccuracies from the dynamic model. The dynamic model increases the amount of predetermined knowledge about how the vessel is likely to move and thus reduces the black-box factor typically experienced in purely data-driven predictors. A prediction horizon of 30 seconds ahead of real time during docking operations is examined. History data from the 29-meter coastal displacement ship RV (Research Vessel) Gunnerus is applied to validate the approach. Results show that the inclusion of the data-based ML model significantly improves the prediction accuracy.

Sea state estimation is a fundamental problem in the development of autonomous ships. Traditional methods such as wave buoy, satellites, and wave radars are limited by locations, clouds and costs, respectively. Model-based methods are prone to incorrect estimations due to their high dependency on mathematical models of ships. As previous data-driven studies for sea state estimation only consider wave height and use the motion data from dynamic positioning vessels, this paper introduces a new, deep neural network (SSENET) to estimate sea state in light of both wave height and wave direction, and extends the generality of sensor data from ship motion with forward speed. SSENET is built on the basis of stacked convolutional neural network blocks with dense connections between different blocks, channel attention modules and a feature attention module. The dense connections build short-cut paths between input and all subsequent convolutional blocks, which can make full use of all the hierarchical features from the original time series sensor data. The channel attention modules aim to enhance the features extracted by each convolution block. The feature attention module focuses on combining the feature fusion of hierarchical features in an adaptive manner. Benchmark experiments show the competitive performance against state-ofthe-art approaches. Applying the SSENET on two datasets of zigzag motion for comparative studies shows the effectiveness of the proposed method.

Maintenance is the key to ensuring the safe and efficient operation of marine vessels. Currently, reactive maintenance and preventive maintenance are two main approaches used onboard. These approaches are either cost-intensive or labor-intensive. Recently, Prognostics and Health Management has emerged as an optimal way to manage maintenance operations. In such a system, fault prognostics aims to predict the remaining useful life based on the sensor measurements. In this paper, the feasibility of applying data-driven fault prognostics to marine diesel engines is investigated. Real-world run-to-failure data of two independent fault-types in two different engine load profiles are collected from a hybrid power lab. The first profile is used for training and validation, while the second is used for testing. The LSTM networks are used to construct the fault prognostics model. Experiments and comparisons are performed to obtain the optimal structure of the networks. Results show that the proposed method generalizes well on the second profile and provides remaining useful life predictions with high accuracy.

Wind farms are usually located in high-latitude areas, which bring a high risk of icing. Traditional methods of anti-blade-icing are limited by extra costs and potential damages to the original mechanical structure. Model-based methods are heavily dependent on mathematical models of the blade icing, which are prone to produce erroneous estimation. As data-driven models are better able to achieve competitive performances for the blade icing estimation, this article proposes a temporal attention-based convolutional neural network (TACNN). This novel data-driven model introduces a temporal attention module into a convolutional neural network, with the goal of determining the importance of sensors and timesteps and automatically identifying discriminative features from raw sensor data. Benchmark experiments on ten public datasets of multivariate time-series classification show competitive performance against the state-of-the-art methods. Compared with ten baseline networks and three widely used attention mechanisms, the TACNN shows significant advantages applying to three real-world datasets. These datasets are logged by the supervisory control and data acquisition system and contain operational and environmental measurements such as power and temperature. The ablation study and sensitivity study demonstrate the effectiveness of the key components of the TACNN. The practicability of the TACNN is further verified through online estimation testing.

This paper presents a data-driven model for time series prediction of ship motion. Prediction based on past time series of data is a powerful function in modern ship support systems. For a large amount of ship sensor data, neural network (NN) is considered as a proper tool in modelling the prediction system. Efforts are made to compact the NN structure through sensitivity analysis, in which the importance of each input to the output is quantified and lower ranked inputs are eliminated. Further analysis about the impact of three different learning strategies, i.e. offline, online and hybrid learning on the NN, is conducted. The hybrid learning combining the advantages of both the offline learning and the online learning exhibits superior prediction performance. According to the long-term prediction ability of recurrent NN, multi-step-ahead prediction under the hybrid learning strategy is realised in a multi-stage prediction form. Experiments are carried out using collected ship sensor data on a vessel. The results show the feasibility of generating a data-driven model through modelling and analysis of the NN for ship motion prediction.